Intent to Experiment: WebNN

[Chromestatus]

Contact emails

dbo...@google.com, rei...@chromium.org, Rafael....@microsoft.com, ningx...@intel.com

Explainer

https://github.com/webmachinelearning/webnn/blob/main/explainer.md

Specification

https://webmachinelearning.github.io/webnn

Design docs

https://docs.google.com/document/d/1kIc5NO1PU1DbGHKOsnan-tEhV01L6_K3OLA5g0o4Ts4/edit?resourcekey=0-ltUDiDZqxKgARhtNqsGJeg

Summary

WebNN seeks to enable web applications and frameworks to take advantage of native operating system services for machine learning and the underlying hardware capabilities available on a user's computer to implement consistent, efficient, and reliable ML experiences on the web.

Blink component

Blink>WebML

Web Feature ID

webnn

TAG review

https://github.com/w3ctag/design-reviews/issues/933

TAG review status

Issues addressed

Origin Trial documentation link

https://webnn.io/en/learn/get-started/quickstart

Risks

Interoperability and Compatibility

Aside from the risk that other browsers do not implement (which is trending towards unlikely given that Edge is planning to ship alongside Chrome and Firefox is planning to implement) the primary interoperability risk is that a WebNN implementation needs to abstract over the ML acceleration capabilities of the underlying platform and hardware. The specification mitigates this risk by defining a set of ML operations which must be supported and provides a feature detection mechanism for discovering additional operations that may be optionally supported. This allows developers to write broadly interoperable code by using only the required operations or design their applications to adapt to additional capabilities at runtime.

Gecko: Positive (https://github.com/mozilla/standards-positions/issues/1215#issuecomment-3520278819)

WebKit: No signal (https://github.com/WebKit/standards-positions/issues/486) Apple is participating in the WebML WG but has made no public statements about implementation.

Web developers: Positive Microsoft's ONNX runtime includes an execution provider for WebNN (https://onnxruntime.ai/docs/tutorials/web/ep-webnn.html). Google's LiteRT.js framework has announced that they are working on WebNN support (https://youtu.be/HAjotVloAvI?t=219).

Other signals: The WebNN implementation in Chromium is a collaboration between Google, Intel, and Microsoft. Edge will be launching WebNN at the same time as Chrome.

Ergonomics

For some use cases, particularly real-time ML processing of video, integration between WebNN and WebGPU is required. Support for this has been prototyped on some platforms as an `exportToGPU()` method that takes an `MLTensor` and makes it available as a `GPUBuffer` but this is not yet available on all platforms.

Activation

Most developers don't write JavaScript or WebGPU shader code to execute their ML models by hand. Instead frameworks such as LiteRT.js, ONNX Runtime Web, WebLLM or Transformers.js are used. Support for WebNN in these frameworks (already starting to appear) will be a large factor in the adoption of this API.

Security

The primary security risk is exposing the underlying platform frameworks to potentially malicious content. This is mitigated through sandboxing of the GPU process and fuzz testing. The risks are similar to WebGPU, though the expressiveness of WebNN is lower than WGSL.

WebView application risks

Does this intent deprecate or change behavior of existing APIs, such that it has potentially high risk for Android WebView-based applications?

This is a new API.

Goals for experimentation

No information provided

Ongoing technical constraints

WebNN can currently execute models on CPU across all platforms but support for GPU and NPU execution is more limited. This will expand as development continues.

Debuggability

There is currently no special DevTools support for this API beyond regular JavaScript debugging. A feature we have considered (and prototyped via an internal extension and the --webnn-coreml-dump-model, --webnn-ort-dump-model, and --webnn-tflite-dump-model browser flags) is inspecting graphs constructed by a site and exporting them in the underlying platform format so that they can be inspected with other tools.

Will this feature be supported on all six Blink platforms (Windows, Mac, Linux, ChromeOS, Android, and Android WebView)?

Yes

Is this feature fully tested by web-platform-tests?

Yes
https://wpt.fyi/results/webnn?label=master&label=experimental&aligned&q=webnn

DevTrial instructions

https://webnn.io/en/learn/get-started/installation

Flag name on about://flags

web-machine-learning-neural-network

Finch feature name

WebMachineLearningNeuralNetwork

Requires code in //chrome?

True

Tracking bug

https://issues.chromium.org/issues/40206287

Estimated milestones

Origin trial desktop first	145
Origin trial desktop last	147
DevTrial on desktop	125
Origin trial Android first	145
Origin trial Android last	147
DevTrial on Android	125
Origin trial WebView first	145
Origin trial WebView last	147

Link to entry on the Chrome Platform Status

https://chromestatus.com/feature/5176273954144256?gate=5186064701194240

Links to previous Intent discussions

Intent to Prototype: https://groups.google.com/a/chromium.org/g/blink-dev/c/PD6TDMDS9mg

This intent message was generated by Chrome Platform Status.

[Reilly Grant]

Due to GoogleChrome/chromium-dashboard#4155 this wasn't filled in. It should read:

The goal of this experiment is to understand how well real-world models abstract to the operations supported by WebNN and how well WebNN can map these operations to those supported by the hardware of real-world users.

[Rick Byers]

Glad to see this in OT, hope to see some cool things in the wild soon! LGTM to experiment 145-147 inclusive. 

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/70b6585e-7685-45d6-9507-dbbf586799ecn%40chromium.org.